November 5, 2021

1290 words 7 mins read



Real-ESRGAN aims at developing Practical Algorithms for General Image Restoration.

repo name xinntao/Real-ESRGAN
repo link
language Python
size (curr.) 3414 kB
stars (curr.) 6375
created 2021-07-19
license BSD 3-Clause “New” or “Revised” License


download PyPI Open issue Closed issue LICENSE python lint Publish-pip

English | 简体中文

  1. Colab Demo for Real-ESRGAN .
  2. Portable Windows / Linux / MacOS executable files for Intel/AMD/Nvidia GPU. You can find more information here. The ncnn implementation is in Real-ESRGAN-ncnn-vulkan.

Thanks for your interests and use:-) There are still many problems about the anime/illustration model, mainly including: 1. It cannot deal with videos; 2. It cannot be aware of depth/depth-of-field; 3. It is not adjustable; 4. May change the original style. Thanks for your valuable feedbacks/suggestions. All the feedbacks are updated in Hopefully, a new model will be available soon.

感谢大家的关注和使用:-) 关于动漫插画的模型,目前还有很多问题,主要有: 1. 视频处理不了; 2. 景深虚化有问题; 3. 不可调节, 效果过了; 4. 改变原来的风格。大家提供了很好的反馈。这些反馈会逐步更新在 这个文档。希望不久之后,有新模型可以使用.

Real-ESRGAN aims at developing Practical Algorithms for General Image Restoration. We extend the powerful ESRGAN to a practical restoration application (namely, Real-ESRGAN), which is trained with pure synthetic data.

:art: Real-ESRGAN needs your contributions. Any contributions are welcome, such as new features/models/typo fixes/suggestions/maintenance, etc. See All contributors are list here.

:question: Frequently Asked Questions can be found in (Well, it is still empty there =-=||).

:triangular_flag_on_post: Updates

  • :white_check_mark: Add the ncnn implementation Real-ESRGAN-ncnn-vulkan.
  • :white_check_mark: Add RealESRGAN_x4plus_anime_6B.pth, which is optimized for anime images with much smaller model size. More details and comparisons with waifu2x are in
  • :white_check_mark: Support finetuning on your own data or paired data (i.e., finetuning ESRGAN). See here
  • :white_check_mark: Integrate GFPGAN to support face enhancement.
  • :white_check_mark: Integrated to Huggingface Spaces with Gradio. See Gradio Web Demo. Thanks @AK391
  • :white_check_mark: Support arbitrary scale with --outscale (It actually further resizes outputs with LANCZOS4). Add RealESRGAN_x2plus.pth model.
  • :white_check_mark: The inference code supports: 1) tile options; 2) images with alpha channel; 3) gray images; 4) 16-bit images.
  • :white_check_mark: The training codes have been released. A detailed guide can be found in

If Real-ESRGAN is helpful in your photos/projects, please help to :star: this repo or recommend it to your friends. Thanks:blush: Other recommended projects: :arrow_forward: GFPGAN: A practical algorithm for real-world face restoration :arrow_forward: BasicSR: An open-source image and video restoration toolbox :arrow_forward: facexlib: A collection that provides useful face-relation functions. :arrow_forward: HandyView: A PyQt5-based image viewer that is handy for view and comparison.

:book: Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data

[Paper]   [Project Page]   [YouTube Video]   [B站讲解]   [Poster]   [PPT slides] Xintao Wang, Liangbin Xie, Chao Dong, Ying Shan Tencent ARC Lab; Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences

We have provided a pretrained model (RealESRGAN_x4plus.pth) with upsampling X4. Note that RealESRGAN may still fail in some cases as the real-world degradations are really too complex. Moreover, it may not perform well on human faces, text, etc, which will be optimized later.

Real-ESRGAN will be a long-term supported project (in my current plan :smiley:). It will be continuously updated in my spare time.

Here is a TODO list in the near future:

  • optimize for human faces
  • optimize for texts
  • optimize for anime images
  • support more scales
  • support controllable restoration strength

If you have any good ideas or demands, please open an issue/discussion to let me know. If you have some images that Real-ESRGAN could not well restored, please also open an issue/discussion. I will record it (but I cannot guarantee to resolve it:stuck_out_tongue:). If necessary, I will open a page to specially record these real-world cases that need to be solved, but the current technology is difficult to handle well.

Portable executable files

You can download Windows / Linux / MacOS executable files for Intel/AMD/Nvidia GPU.

This executable file is portable and includes all the binaries and models required. No CUDA or PyTorch environment is needed.

You can simply run the following command (the Windows example, more information is in the of each executable files):

./realesrgan-ncnn-vulkan.exe -i input.jpg -o output.png

We have provided three models:

  1. realesrgan-x4plus (default)
  2. realesrnet-x4plus
  3. realesrgan-x4plus-anime (optimized for anime images, small model size)

You can use the -n argument for other models, for example, ./realesrgan-ncnn-vulkan.exe -i input.jpg -o output.png -n realesrnet-x4plus

Usage of executable files

  1. Please refer to Real-ESRGAN-ncnn-vulkan for more details.
  2. Note that it does not support all the functions (such as outscale) as the python script
Usage: realesrgan-ncnn-vulkan.exe -i infile -o outfile [options]...

  -h                   show this help
  -v                   verbose output
  -i input-path        input image path (jpg/png/webp) or directory
  -o output-path       output image path (jpg/png/webp) or directory
  -s scale             upscale ratio (4, default=4)
  -t tile-size         tile size (>=32/0=auto, default=0) can be 0,0,0 for multi-gpu
  -m model-path        folder path to pre-trained models(default=models)
  -n model-name        model name (default=realesrgan-x4plus, can be realesrgan-x4plus | realesrgan-x4plus-anime | realesrnet-x4plus)
  -g gpu-id            gpu device to use (default=0) can be 0,1,2 for multi-gpu
  -j load:proc:save    thread count for load/proc/save (default=1:2:2) can be 1:2,2,2:2 for multi-gpu
  -x                   enable tta mode
  -f format            output image format (jpg/png/webp, default=ext/png)

Note that it may introduce block inconsistency (and also generate slightly different results from the PyTorch implementation), because this executable file first crops the input image into several tiles, and then processes them separately, finally stitches together.

This executable file is based on the wonderful Tencent/ncnn and realsr-ncnn-vulkan by nihui.

:wrench: Dependencies and Installation


  1. Clone repo

    git clone
    cd Real-ESRGAN
  2. Install dependent packages

    # Install basicsr -
    # We use BasicSR for both training and inference
    pip install basicsr
    # facexlib and gfpgan are for face enhancement
    pip install facexlib
    pip install gfpgan
    pip install -r requirements.txt
    python develop

:zap: Quick Inference

Inference general images

Download pre-trained models: RealESRGAN_x4plus.pth

wget -P experiments/pretrained_models


python --model_path experiments/pretrained_models/RealESRGAN_x4plus.pth --input inputs --face_enhance

Results are in the results folder

Inference anime images

Pre-trained models: RealESRGAN_x4plus_anime_6B More details and comparisons with waifu2x are in

# download model
wget -P experiments/pretrained_models
# inference
python --model_path experiments/pretrained_models/RealESRGAN_x4plus_anime_6B.pth --input inputs

Results are in the results folder

Usage of python script

  1. You can use X4 model for arbitrary output size with the argument outscale. The program will further perform cheap resize operation after the Real-ESRGAN output.
Usage: python --model_path experiments/pretrained_models/RealESRGAN_x4plus.pth --input infile --output outfile [options]...

A common command: python --model_path experiments/pretrained_models/RealESRGAN_x4plus.pth --input infile --netscale 4 --outscale 3.5 --half --face_enhance

  -h                   show this help
  --input              Input image or folder. Default: inputs
  --output             Output folder. Default: results
  --model_path         Path to the pre-trained model. Default: experiments/pretrained_models/RealESRGAN_x4plus.pth
  --netscale           Upsample scale factor of the network. Default: 4
  --outscale           The final upsampling scale of the image. Default: 4
  --suffix             Suffix of the restored image. Default: out
  --tile               Tile size, 0 for no tile during testing. Default: 0
  --face_enhance       Whether to use GFPGAN to enhance face. Default: False
  --half               Whether to use half precision during inference. Default: False
  --ext                Image extension. Options: auto | jpg | png, auto means using the same extension as inputs. Default: auto

:european_castle: Model Zoo

The following models are discriminators, which are usually used for fine-tuning.

:computer: Training and Finetuning on your own dataset

A detailed guide can be found in


    author    = {Xintao Wang and Liangbin Xie and Chao Dong and Ying Shan},
    title     = {Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data},
    booktitle = {International Conference on Computer Vision Workshops (ICCVW)},
    date      = {2021}

:e-mail: Contact

If you have any question, please email or

:hugs: Acknowledgement

Thanks for all the contributors.

comments powered by Disqus